perf: use trunc(dval(&u)) instead of (Long)(dval(&u))

[henderkes]

… to keep per-digit dependency chain in float register

<?php
// Micro-benchmark of the float->string paths
const N = 4000;
const R = 8;

$vals = [];
for ($i = 1; $i <= N; $i++) {
    $vals[] = M_PI * $i / 7.0 + 1.0 / $i;
}

function bench(string $name, callable $fn, array $vals): void {
    $fn($vals); $fn($vals);
    $best = PHP_INT_MAX;
    for ($r = 0; $r < R; $r++) {
        $t = hrtime(true);
        $sink = $fn($vals);
        $dt = hrtime(true) - $t;
        if ($dt < $best) $best = $dt;
    }
    $ops = count($vals);
    printf("%-26s %8.1f ns/op   (min block %6.2f ms)\n", $name, $best / $ops, $best / 1e6);
}

bench('(string) cast', function($vals){ $s=0; foreach($vals as $v){ $x=(string)$v; $s+=strlen($x); } return $s; }, $vals);
bench('string interpolation', function($vals){ $s=0; foreach($vals as $v){ $x="$v"; $s+=strlen($x); } return $s; }, $vals);
bench('concatenation', function($vals){ $s=''; foreach($vals as $v){ $s = $v . '|'; } return strlen($s); }, $vals);
bench('number_format', function($vals){ $s=0; foreach($vals as $v){ $s+=strlen(number_format($v, 6, '.', ',')); } return $s; }, $vals);
bench("sprintf('%g')", function($vals){ $s=0; foreach($vals as $v){ $s+=strlen(sprintf('%g',$v)); } return $s; }, $vals);
bench("sprintf('%.10f')", function($vals){ $s=0; foreach($vals as $v){ $s+=strlen(sprintf('%.10f',$v)); } return $s; }, $vals);
bench('json_encode (array)', function($vals){ return strlen(json_encode($vals)); }, $vals);
bench('serialize (array)', function($vals){ return strlen(serialize($vals)); }, $vals);
bench('var_export (array)', function($vals){ return strlen(var_export($vals, true)); }, $vals);
bench('implode floats', function($vals){ return strlen(implode(',', $vals)); }, $vals);

aarch64:

Path (ns/op)	base	dtoa	change
(string) cast	182.6	146.4	−19.8%
string interpolation	182.8	146.3	−20.0%
concatenation	196.2	159.5	−18.7%
number_format	309.8	286.4	−7.6%
sprintf('%g')	153.0	137.6	−10.1%
sprintf('%.10f')	258.8	226.4	−12.5%
json_encode (array)	568.1	567.8	−0.1%
serialize (array)	580.9	579.7	−0.2%
var_export (array)	637.5	635.9	−0.3%
implode floats	164.9	129.3	−21.6%

x86_64 (sse4.1):

Path (ns/op)	base	dtoa	change
(string) cast	105.9	100.1	−5.5%
string interpolation	105.6	100.3	−5.0%
concatenation	111.5	106.1	−4.8%
number_format	162.4	158.8	−2.2%
sprintf('%g')	82.2	79.8	−2.9%
sprintf('%.10f')	136.9	130.7	−4.5%
json_encode (array)	261.1	259.7	−0.5%
serialize (array)	268.2	266.8	−0.5%
var_export (array)	287.5	290.1	+0.9%
implode floats	103.1	97.6	−5.3%

LLM disclosure: optimisation found entirely by Fable, I verified, wrote the comment and benchmarked.

[Girgias]

This file is old and complicated and previous changees were rejected.

IMHO it would make more sense to replace this with a known correct and fast alternative like https://github.com/kolemannix/ffc.h

[henderkes]

I'm sure replacing the whole logic would be a good idea, but I don't see why the file being old should lead to changes being rejected. It's a rather minimal change with decent performance gains (especially on aarch64) for a common operation and doesn't make replacing the logic entirely in the future any harder.